Prediction of soil clay minerals from some soil properties with use of feature selection algorithm and ANFIS methods

Abstract

Researchers use different methods to investigate and quantify clay minerals. X-ray diffraction is a common and widespread approach for clay mineralogy investigation, but is time-consuming and expensive, especially in highly calcareous soils. The aim of this research was prediction of clay minerals in calcareous soils of southern Iran using a feature selection algorithm and adaptive neuro-fuzzy inference system (ANFIS) methods. Fifty soil samples from different climatic regions of southern Iran were collected and different climatic, soil properties and clay minerals were determined using X-ray diffraction. Feature selection algorithms were used for selection of the best feature subset for prediction of clay mineral types along with two sets of training and testing data. Results indicated that the best feature subset by Best-First for prediction of illite was cation exchange capacity (CEC), sand, total potassium, silt and agroclimatic index (correlation coefficient (R) = 0.99 for training and testing data); for smectite was precipitation, temperature, evapotranspiration and CEC (R = 0.89 and 0.87 for training and testing data respectively); and for palygorskite was precipitation, temperature, evapotranspiration and calcium carbonate equivalent (CCE) (R = 0.98 for training and testing data). An attempt was made to predict clay minerals type by ANFIS using selected data from the feature selection algorithm. The evaluation of method by calculating root mean square error (RMSE), mean absolute error (MAE) and R indicated that the ANFIS method may be suitable for illite, chlorite, smectite and palygorskite prediction (RMSE, MAE and R of 0.001–0.028, 0.004–0.012 and 0.67–0.89 respectively for training and testing data). Comparison of data for all clay minerals showed that ANFIS method did not predict illite and chlorite as well as other minerals in the studied soils.